What are reasoning or thinking models in ChatGPT?

Reasoning models, often shown in ChatGPT as a thinking mode, are AI models trained to work through a problem step by step before giving a final answer. Instead of replying immediately, they spend extra time "thinking," exploring approaches, checking their work and correcting mistakes.

OpenAI introduced its first reasoning models in 2024, and other companies followed with similar approaches, such as Claude's extended thinking, Gemini's thinking models and DeepSeek-R1. The idea is now a standard part of modern AI.

How does it work? A regular chat model generates its answer directly, word by word. A reasoning model first produces an internal chain of reasoning, working through the problem much as a student writes rough work in the margin before the final answer. It was trained specifically to reason this way, being rewarded for reaching correct answers on difficult problems. You often see a "thinking" indicator, and sometimes a summary of the reasoning.

When are thinking models better?

1. Maths and logic problems, such as multi-step calculations or puzzles.
2. Coding tasks with complex logic or debugging.
3. Science and engineering questions.
4. Analysing a complicated situation with many factors, such as a business decision.
5. Planning tasks with multiple constraints, such as a timetable or project schedule.
6. Questions where accuracy matters more than speed.

When is a regular model enough? For simple questions, quick facts, casual conversation, short emails and creative writing, a regular fast model is usually just as good and much quicker. Thinking models are slower, sometimes taking from several seconds to a few minutes, and on many plans they have tighter usage limits.

In current versions of ChatGPT, you can often choose a thinking option from the model menu, or ChatGPT may automatically decide when to think longer based on your question. You can also nudge it by asking it to "think carefully" about a hard problem.

A few tips help. Give complete information upfront, because reasoning models do their best work when the problem is clearly defined. Do not over-explain how to reason, since these models already have their own approach. And still verify important results, because extra thinking reduces mistakes but does not eliminate them.

Think of it like the difference between answering a question off the top of your head and sitting down with a notebook. Both have their place.